A machine learning model for high-throughput analysis of RT / S region sequence features of the hepatitis B virus genome to predict liver cancer risk

Abstract

The present invention relates to a method for predicting liver cancer risk by using a machine learning model to analyze the sequence characteristics of the RT / S region of the hepatitis B virus genome with high throughput, and belongs to the field of genome sequence calculation technology. The present invention selects a corresponding feature calculation method to complete the construction of the HBV genome RT / S region sequence feature matrix; then uses a machine learning model to train the data set, and finally predicts the probability of liver cancer risk. Based on an artificial intelligence method, the present invention applies a series of feature algorithms to the learning of machine learning model training in the field of HBV genome RT / S region sequence data processing, solves the problem of converting high-dimensional genome data sequencing signals into clinically applicable disease risk judgment, and helps computer-assisted HBV-infected people to have a high-risk early warning and early diagnosis of liver cancer.

Description

Technical Field

[0001] The present invention relates to a method for predicting liver cancer risk by using a machine learning model to perform high-throughput analysis of sequence features in the RT / S region of the hepatitis B virus (HBV) genome, belonging to the technical field of genome sequence calculation, and specifically to a method for calculating sequence features in the RT / S region of the HBV genome and training and predicting a machine learning model, which is used for high-risk warning and early diagnosis of liver cancer in HBV-infected people and has clinical application value. Background Art

[0002] Hepatocellular carcinoma (HCC) is the most common primary liver cancer, the third leading cause of cancer death, and the sixth most common cancer in the world. HBV infection is one of the most important risk factors. According to the World Health Organization (WHO), approximately 257 million people worldwide are confirmed to be positive for hepatitis B virus surface antigen (HBsAg). Chronic hepatitis B (CHB) patients will gradually develop cirrhosis or even liver cancer if not treated in time.

[0003] The HBV genome is a circular double-stranded DNA molecule with some single-stranded regions, consisting of four overlapping open reading frames (ORFs): P, C, S, and X regions. The HBV reverse transcriptase (RT) gene in the P-ORF encodes the RT region of the polymerase protein, which partially overlaps with the S gene encoding HBsAg and is an important region for antiviral resistance and immune escape. Mutations in the HBV genome are closely associated with HCC, mainly due to the high replication rate of HBV, host immune pressure, and error-prone HBV reverse transcriptase.

[0004] Currently, most studies are based on Sanger sequencing to detect HBV gene mutations. However, due to its low sensitivity (about 20%), traditional Sanger sequencing has difficulty in identifying smaller viral mutations and characterizing the complex patterns of HBV quasispecies in a single patient. Compared with first-generation sequencing, next generation sequencing (NGS) can obtain a large number of HBV quasispecies sequence fragments while maintaining high sensitivity and high specificity. It can accurately and deeply explore HBV mutation information, and further extract sequence features based on a large amount of sequence information for bioinformatics analysis. However, to date, few studies have analyzed and studied the HBV RT gene mutation characteristics of CHB and HBV-related HCC patients using big data generated by NGS technology. In addition, traditional statistical algorithms seem to be insufficient to process such high-dimensional genomic data and cannot convert quality sequencing signals into feasible biomarkers for clinical use.

[0005] In response to the above problems, the present invention performs feature calculation on the HBV genome RT / S region sequence data generated by NGS sequencing, and integrates a machine learning model to identify the internal quasispecies characteristics of the HBV RT / S region to predict individual HCC risk. This method has important clinical application prospects in HCC risk screening, auxiliary clinical diagnosis and disease stratification management, and is the first calculation method for HCC risk prediction using dynamic HBV quasispecies characteristics and machine learning models. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a method for predicting liver cancer risk based on HBV genome RT / S region sequence features for a machine learning model in view of the deficiencies of existing NGS technology in processing high-dimensional HBV genome data and effectively converting it into a liver cancer risk model that can be used clinically. The feature matrix generated by the algorithm is applied to the training and learning of the machine learning model for genome sequence data processing, and by establishing a classification model, the identification of CHB (non-HCC) and HCC populations is achieved, so that it can play a role in the auxiliary diagnosis of HCC and clinical high-risk warning.

[0007] In order to solve the above technical problems, the present invention provides a method for predicting liver cancer risk by high-throughput analysis of HBV genome RT / S region sequence characteristics using a machine learning model, comprising the following steps:

[0008] (1) Data preprocessing: For sequence data, determine whether the sequence storage type is FASTA format, determine whether it is a reverse complementary sequence, and select the corresponding method for preprocessing based on the judgment results;

[0009] (2) Generate the amino acid sequences of RT region and S region;

[0010] (3) Construct feature matrices for the RT region nucleotide sequence and the RT region / S region amino acid sequence respectively;

[0011] (4) Divide the feature matrix data set into a training set and a test set, input the training set into the machine learning model for parameter training, and perform probability prediction.

[0012] (5) Use the trained model to evaluate the robustness and various performance of the model, such as sensitivity, specificity, AUC value, accuracy, etc., on a new batch of patient data.

[0013] In the step (2), the translation of the amino acid sequence of the RT region starts from the first nucleotide, and the translation of the amino acid sequence of the S region starts from the second nucleotide, using the reference sequence (https: / / www.ncbi.nlm.nih.gov / projects / genotyping / view.cgi?db=2).

[0014] In the step (3), the feature algorithms include ten types: base distribution percentage feature of the site (ACGT feature), mutation frequency feature of RT region nucleotide sequence (Mutation feature), mutation frequency feature of RT region amino acid sequence (Mutation feature of HBV RT amino acid), mutation frequency feature of S region amino acid sequence (Mutation feature of HBsAg amino acid), Shannon entropy feature of RT region nucleotide sequence (Entropy feature), Shannon entropy feature of RT region amino acid sequence (Entropy feature of HBV rt amino acid), Shannon entropy feature of S region amino acid sequence (Entropy feature of HBsAg amino acid), Group Walk entropy feature of RT region nucleotide sequence, Slide entropy feature of RT region nucleotide sequence and 3-mer feature (K-mer feature) of RT region sequence.

[0015] In the step (4), the machine learning training models are four types: Random Forests (RF), k-Nearest Neighbor (kNN), Support Vector Machine (SVM) and Logistic Regression (LR). The machine learning models used have good applicability. More typically, in the classification model RF, classification for different HBV genotypes and sequencing depths can effectively improve the prediction accuracy.

[0016] In this method, the RT / S region sequence of the HBV genome data is divided into two fragments, the RT1 sequence length is 481 bp, and the RT2 sequence length is 406 bp.

[0017] In this method, for each trained model, a total of 11 different features (10 single features and 1 combination of all features) were input, and a total of 44 different models were finally obtained.

[0018] The feature matrix of the method also adds all features, including the integration of ten features, to improve the prediction ability of the model. After the model training and testing are completed, the model with the best area under the curve AUC is selected as the model for distinguishing CHB and HCC populations, and the model output value is used as the HCC risk prediction score. The established model also has a good area under the curve after independent sample verification.

[0019] The machine learning model based on artificial intelligence in the present invention applies the sequence feature algorithm to the machine learning model training in the field of NGS sequencing data processing, which can effectively solve the processing and clinical application of NGS high-dimensional sequence data, and help to use machine learning to perform computer-assisted screening of high-risk HCC populations, thereby improving the detection rate and clinical precision treatment of the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is an AUC evaluation graph of sample classification trained and predicted using a machine learning method according to an embodiment of the present invention.

[0021] Figure 2 This is a comparison chart of the classification performance of the model used in the embodiment of the present invention for two HBV genotype populations.

[0022] Figure 3 This is a comparison chart of model classification performance for different sequencing depths according to an embodiment of the present invention.

[0023] Figure 4 It is a flow chart of the risk prediction of hepatitis B-related liver cancer of the present invention. DETAILED DESCRIPTION

[0024] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0025] Example: Model discrimination analysis of a multicenter cohort of CHB and HCC patients

[0026] A total of 237 HCC patients and 307 multicenter CHB patients who met the inclusion criteria were collected from Oriental Hepatobiliary Surgery Hospital from 2015 to 2019. All included patients met the following conditions: serum HBsAg was continuously positive for more than 6 months, serum HBV DNA was greater than 10 4 IU / ml, all patients underwent radical liver cancer resection, the tumor was completely removed with sufficient safety margins, and all patients were finally confirmed to have primary hepatocellular carcinoma by histopathological examination. Exclusion criteria include: patients who have received liver transplantation in the past, suffered from other malignant tumors, tumors of unknown origin, metastatic hepatocellular carcinoma, autoimmune liver disease, drug-induced liver disease, alcoholic hepatitis or other chronic liver diseases (such as hepatitis C or D, etc.), and combined human immunodeficiency virus infection. The tumor resection specimens of the selected patients were pathologically confirmed by three pathology experts to ensure negative margins, and the tumor characteristics were recorded in detail: including the maximum diameter of the tumor, the number of tumors, whether the capsule is intact, whether there are satellite lesions, whether it is metastatic, whether there is portal vein cancer thrombus, etc., and TNM grading was performed. Pericancerous tissue is defined as tissue within 3-5 cm from the tumor margin. All tissue samples were immediately frozen at -80℃ for future use after resection. Inclusion criteria for CHB patients: patients with HBsAg positivity for more than 6 months, repeated or persistent elevation of serum ALT (more than 2 times the upper limit of the reference range, and excluding other causes of ALT elevation), or liver pathology with features of chronic viral hepatitis, detectable HBV DNA (>10 4 IU / ml) can be diagnosed as CHB. Exclusion criteria: liver cancer, tumors of other organs or major diseases are excluded. All selected participants signed informed consent. 200μl serum samples were taken from the above personnel for NGS detection of HBV genome RT region.

[0027] like Figure 1 As shown, the method for predicting liver cancer risk by using a machine learning model to analyze the sequence characteristics of the RT / S region of the HBV genome with high throughput comprises the following steps:

[0028] 1) Data preprocessing;

[0029] a. Perform fasta conversion on fastq sequence data,

[0030] b. Determine whether the sequence is a reverse complementary sequence. If so, perform reverse complementary conversion.

[0031] c. Convert the sequence A, T, C, G to the numbers 0, 1, 2, 3.

[0032] 2) Generate RT region and S region amino acid sequences:

[0033] a. Translate the nucleotide sequences of the sequencing results RT1 and RT2 regions into the protein sequences of the RT region. The translation start-stop sites are: 2 th nt-end,

[0034] b. Translate the nucleotide sequences of the RT1 and RT2 regions into the protein sequence of the S region. The translation start-stop sites of the RT1 region are: 3 th nt-479 th nt (the third from the end of the segment), 41AA-199AA (a total of 159aa); the translation start-stop site in the RT2 region is: 3 th nt-104 th nt, 193aa-226aa (total 34aa).

[0035] 3) Feature matrix construction:

[0036] a. For the jth sequence of the i-th sample, define S ij =(n ij1 , n ij2 ,...n ij1 ), where n ijk ∈∑ (∑ is the nucleotide character set, ∑ = [A, C, G, T]), 1 represents the sequence length (for RT1 sequence, set 1 = 481, for RT2 sequence, set 1 = 406), for each sample, use depth i To describe the sequencing depth of the i-th sample. The sequence of each sample is represented by S i =(S i1 , S i2 , ..., S i，depthi );

[0037] b. Set the positions of the ten features in the RT1 region in the feature matrix as ACGT (1-1924), Mutation (1925-2405), Entropy (2406-2886), Group Walk (2887-3046), Slide (3047-3525), RT_entropy (3526-3685), S_entropy (3686-3844), RT_mutation (3845-4004), S_mutation (4005-4163) and K-mer (4164-4227), and set the 11th feature (all features) as the integration of the above ten features, with a position of 1-4227 in the feature matrix;

[0038] c. Set the positions of the ten features in the RT2 region in the feature matrix as ACGT (1-1624), Mutation (1625-2030), Entropy (2437-2571), Group Walk (2572-2975), Slide (2572-2975), RT_entropy (2976-3110), S_entropy (3111-3144), RT_mutation (3145-3279), S_mutation (3280-3313) and K-mer (3314-3377). Set the 11th feature (all features) as the integration of the above ten features, and its position in the feature matrix is ​​1-3377.

[0039] d. Based on the above information, select the corresponding feature algorithm for calculation.

[0040] 4) Machine learning model training and learning

[0041] a. Based on the above feature matrix, the above samples are divided into two cohorts, including a training set (HCC = 159, CHB = 202) and a test set (HCC = 78, CHB = 105). Define (X_i, Y_i) to represent the training data, where X_i is the extracted feature and Y_i∈[0.1] is the label indicating whether the sample belongs to HCC (Y_i = 0) or CHB (Y_i = 1);

[0042] b. Based on the nonlinear function f(X_i) of the machine learning method, calculate the probability P(Y_i=1|X_i,θ);

[0043] c. Four different machine learning models were used, including: Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbor (KNN) and Logistic Regression (LR). For each model, cross-validation was performed in the training data to optimize the parameters, and finally the optimal parameters were selected for model construction.

[0044] 5) Using the trained models, use independent sample cohorts to evaluate the performance of each model

[0045] a. The AUC value of each model is calculated by ROC curve to compare the prediction performance differences of the four machine learning models under different features;

[0046] b. Divide the data into two groups: HBV genotype B and genotype C. Compare the prediction performance differences of the four machine learning models in genotype B and C data based on the “all features” feature;

[0047] c. 450 samples with more than 5000 reads were selected from 544 samples, and then the data set was randomly divided into two groups, including a training set (CHB = 179; HCC = 125) and an independent test set (CHB = 91; HCC = 55). N reads were randomly selected for each sample, and the relationship between the size of the sample read number and the prediction accuracy of the four models was compared based on the "all features" feature. Among them, N ranged from 500 to 5000, with a step size of 500.

[0048] The identification result of the present invention:

[0049] Depend on Figure 1 It can be seen that for the HBV RT1 region, the AUC values ​​of the four machine learning models using the six features "entropy", "group", "slides", "RT_entropy", "S_entropy" and "all features" are all around 0.8, while the AUCs of the other features in the SVM, KNN, and LR models are all lower than 0.8. Except for the "kmer" feature, the AUCs of the remaining features in the RF model are all stable at around 0.9, while the AUC values ​​of the SVM, KNN, and LR models vary greatly depending on the features used. The RF model showed high accuracy and AUC values ​​in both training and validation data, and the performance of the model was relatively stable under different features. The results for the HBV RT2 region were consistent with those for the RT1 region.

[0050] Depend on Figure 2It can be seen that for the HBV RT1 and RT2 regions, the four models based on "all features" have roughly the same prediction performance in genotype B and C data. The RF model shows a more stable ability to predict genotype B and genotype C sample data, while the SVM and LR models have better prediction effects on genotype C than on genotype B. In the training set and validation set, the AUC value of the RF model for the prediction of the two HBV genotypes almost reached 0.9. This shows that the RF model has stable prediction performance not only in different HBV genome regions (RT1 and RT2), different feature algorithms, and different HBV genotypes (B and C).

[0051] Depend on Figure 3 It can be seen that in the training set and validation set data, the AUC value of the RF model is stable at around 0.9, the SVM and LR models are roughly maintained at 0.8 in both sets of data, and the AUC value of the KNN model fluctuates between 0.6-0.7. The results show that the performance of the RF model is better than the other three machine learning models. When the number of sample reads is greater than 500, its AUC value is relatively stable.

[0052] Table 1: Basic clinical information of chronic hepatitis B (CHB) and hepatocellular carcinoma (HCC) population

[0053]

[0054] Table 1 shows the basic clinical information of the CHB and HCC populations. There were no significant differences in age, gender, HBV DNA level, and genotype (%) of CHB patients between the training group and the validation group (P values ​​> 0.05). There were no differences in the baseline data of HCC patients between the training group and the validation group (all p values ​​> 0.05). Among them, genotype C accounted for nearly 70% of CHB and HCC in the two cohorts, respectively. In addition, genotype B accounted for almost 27% of CHB in the training group and the validation group, while genotype B accounted for about 14% of HCC patients in both groups. No significant differences were observed for clinical indicators of other liver diseases such as ALT, AST, and GGT (P values ​​> 0.05). In summary, the data showed that there were no statistically significant differences in the baseline information between CHB / HCC patients between the training group and the validation group.

[0055] Table 2 Prediction evaluation table of four machine learning models under different features in training set and validation set

[0056]

[0057] Table 2 shows the evaluation results of the four models in the training set and validation set based on the "all features" feature. Among them, the AUC of the RF model in HBV RT1 and RT2 are 0.9585 and 0.9396 respectively, and its accuracy is also the highest among the four models, reaching 0.9016 and 0.8798 in HBV RT1 and RT2 respectively.

[0058] The technical contents of the present invention are described above in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the said contents. Within the knowledge scope of ordinary technicians in this field, various changes can be made to the technical contents of the present invention without departing from the purpose of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting liver cancer risk by high-throughput analysis of RT / S region sequence features of the hepatitis B virus genome using a machine learning model. Its characteristics include the following steps: (1) Data input and preprocessing: Input the HBV genome RT / S region sequence data, determine whether the file storage type is in FASTA format, determine whether it is a reverse complementary sequence, and select the corresponding method for preprocessing based on the judgment results; the input data is the sequence data of NGS high-throughput sequencing, including two fragments of HBV genome data RT and S region sequences, of which the RT1 sequence length is 481 bp and the RT2 sequence length is 406 bp, and the file format is FASTA or FASTQ; (2) generating the RT / S region amino acid sequence; (3) Feature matrix construction; the feature algorithms are the base distribution percentage feature of the site, the mutation frequency feature of the nucleotide sequence in the RT region, the mutation frequency feature of the amino acid sequence in the RT region, the mutation frequency feature of the amino acid sequence in the S region, the Shannon entropy feature of the nucleotide sequence in the RT region, the Shannon entropy feature of the amino acid sequence in the RT region, the Shannon entropy feature of the amino acid sequence in the S region, the walking Shannon entropy feature of the nucleotide sequence in the RT region, the sliding Shannon entropy feature of the nucleotide sequence in the RT region, and the 3-mer feature of the nucleotide sequence in the RT region. The feature matrix introduces the allfeatures feature to integrate the above features; (4) Dividing the feature matrix data set into a training set and a test set, inputting the training set into the machine learning model for parameter training and probability prediction; the input data is the feature matrix of the HBV genome RT / S region sequence, and four machine learning models, random forest, K nearest neighbor, support vector machine and logistic regression, are used for learning and training; in this step, the machine learning algorithm uses the scikit-learn package; (5) Use the trained model to evaluate the sensitivity, specificity, AUC value, and accuracy of the model on a new batch of patient data; For each trained model, 10 single features and 1 integrated feature of all features were input, and finally 44 different models were obtained. After the model training and testing were completed, the model with the best area under the curve (AUC) was selected as the model for distinguishing chronic hepatitis B and HCC populations, and the model output value was used as the liver cancer risk prediction score. The model with the best area under the curve AUC is the random forest model based on all features.

Images